Qwen Image 2.1 LoRA · Text to Image

The 2.1 family splits along an awkward line. The LoRA variants are cheap and flexible - from about 4.5 credits ($0.045) an image here - but they drop the transparency controls and the prompt-rewriter toggle the plain 2.1 node has. If what you want is "2.1 quality, my own style, no surprises from an automatic rewrite", that trade is the right one.

What the schema gives you

Five widgets and one socket. prompt, aspect_ratio (1:1 default), resolution (1k tier default, 2K available), output_format (JPEG default), seed - and loras, which is the reason to pick this node over its plain sibling.

Worth noticing what's not there, because it's the same JSON schema mirrored into widgets and not a GUI limitation: no prompt-expansion toggle, no background/transparency switch, no multi-output count. One prompt, one image, no rewriter second-guessing your wording. A lot of people prefer it for exactly that.

Outputs are images and task id - no alpha mask, so if you need a cutout go to Qwen Image 2.1 · Text to Image, which exposes one.

Attaching a LoRA

Add a PopNow LoRA node and connect it to loras. Each node carries a source and a scale:

• Local file - a .safetensors in ComfyUI/models/loras. Uploaded once (2 GiB ceiling) and cached for twelve hours, so a batch run doesn't re-upload it thirty times.
• Hugging Face or Civitai - pass the repo id or page URL; the service imports it once. Gated downloads read HF_TOKEN or CIVITAI_API_KEY from the environment, never from a widget.
• A direct https:// link - used as-is.

Stack up to three. My usual starting point: style LoRA 0.7, character LoRA 0.6, and only reach for a third if there's a hole. Two at full strength fight; one at 1.0 tends to flatten everything else in the prompt.

Match the LoRA to the family, too. A 2512-trained LoRA on a 2.1 base will load and render, and "render" is all you can confidently say about it.

Why 2.1 rather than an older revision

Better instruction following than 1.0 and 2512, a 2K tier, and the same 20B Apache-2.0 lineage underneath - this is the same model family the KB describes as the open editing standard, just on a newer checkpoint. If your LoRA is 2.1-trained, there's no reason to run anything older.

Prompting note that holds across the family: the encoder is an LLM, so a descriptive paragraph works and keyword weighting is dead weight. Put a trigger word at the front - the prompt is read in order and a rare token lands cleanest there.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart. Or Manager → Custom Nodes Manager → PopNow → Install, or comfy node install popnow-comfyui. No dependency step exists: the pack's own metadata lists zero dependencies, because it uses the aiohttp, torch, Pillow and NumPy ComfyUI already has. You need Python 3.10+ and a ComfyUI with the V3 node API.

Then the key: popnow.ai/api/console/keys, Settings → PopNow, paste, restart. Once more than you'd think - the node list comes from your account's catalog at startup.

Small print that saves money

seed defaults to fixed. Queue the same graph twice and the second run returns the cached image rather than charging again - genuinely nice, occasionally confusing. Control after generate → randomize for a new take, and remember that every randomize is a charge.

Set Settings → PopNow → Max credits per run before batching. Each node quotes the run first and stops if the quote exceeds your ceiling, so the ceiling is a real guard, not a UI nicety.

Errors are worth reading rather than googling: 40201 no credits, 400 names the field it rejected, 40004 means that combination isn't servable, 40310 wants your email verified. Failed runs are refunded, and task id is what you hand support when something ends in request_id.